Cognitive load in the multi-player prisoner’s dilemma game · Munich Personal RePEc Archive...
Transcript of Cognitive load in the multi-player prisoner’s dilemma game · Munich Personal RePEc Archive...
Munich Personal RePEc Archive
Cognitive load in the multi-player
prisoner’s dilemma game
Duffy, Sean and Smith, John
Rutgers University-Camden
11 May 2011
Online at https://mpra.ub.uni-muenchen.de/30856/
MPRA Paper No. 30856, posted 11 May 2011 12:36 UTC
Cognitive Load in the Multi-player Prisoner�s Dilemma Game:
Are There Brains in Games?�
Sean Du¤yy and John Smithz
May 11, 2011
Abstract
We �nd that di¤erences in the ability to devote cognitive resources to a strategic in-teraction imply di¤erences in strategic behavior. In our experiment, we manipulate theavailability of cognitive resources by applying a di¤erential cognitive load. In cognitiveload experiments, subjects are directed to perform a task which occupies cognitive re-sources, in addition to making a choice in another domain. The greater the cognitiveresources required for the task implies that fewer such resources will be available for delib-eration on the choice. Although much is known about how subjects make decisions undera cognitive load, little is known about how this a¤ects behavior in strategic games. Werun an experiment in which subjects play a repeated multi-player prisoner�s dilemma gameunder two cognitive load treatments. In one treatment, subjects are placed under a highcognitive load (given a 7 digit number to recall) and subjects in the other are placed undera low cognitive load (given a 2 digit number). We �nd that the individual behavior ofthe subjects in the low load condition converges to the Subgame Perfect Nash Equilibriumprediction at a faster rate than those in the high load treatment. However, we do not�nd the corresponding relationship involving outcomes in the game. Speci�cally, there isno evidence of a signi�cantly di¤erent convergence of game outcomes across treatments.As an explanation of these two results, we �nd evidence that low load subjects are betterable to adjust their choice in response to outcomes in previous periods.
**Preliminary and incomplete****Suggestions welcome**
Keywords: cognitive resources, experimental economics, experimental game theory, publicgoods game
JEL: C72, C91
�We wish to thank Hans Czap, Natalia Czap, Tyson Hartwig, Hrvoje Stojic, Roel Van Veldhuizer, JackWorrall, and participants at the 7th IMEBE in Barcelona for helpful comments. This research was supportedby Rutgers University Research Council Grant #202171.
yRutgers University-Camden, Department of Psychology, 311 N. 5th Street, Camden, New Jersey, USA,08102.
zCorresponding Author; Rutgers University-Camden, Department of Economics, 311 North 5th Street,Camden, New Jersey, USA 08102; Email: [email protected]; Phone: +1 856 225-6319; Fax: +1 856225-6602.
1
1 Introduction
There have been advancements in the understanding of play in games based on the conceptu-
alization that players devote heterogenous levels of cognition to deliberation on their strategy
(Stahl and Wilson, 1994, 1995; Nagel, 1995; Costa-Gomes et al., 2001; Camerer et al., 2004).
These advancements specify that the players exhibit heterogenous levels of strategic sophisti-
cation. In particular, it is conceptualized that higher levels within the hierarchy are associated
with greater sophistication. This conceptualization is often supported by observing play in
a game and determining whether the hierarchical model improves the �t with the observa-
tions. In addition to comparing the predictions with the observations, these models are also
supported by the measurement of data related to the level of cognition. For instance stud-
ies measuring the decision to lookup relevant and available information,1 eyetracking studies
which measure the location of the attention of the subject,2 and even neurological data3 have
been seen as providing evidence in support of these hierarchical models.
In a rough sense, these papers ask the questions, "Are there brains in games?" and "If so,
what else can we say?" In our paper, rather than measure the level of cognition or measure
data related to the level cognition, we manipulate the level of cognition. In this sense, the
present paper is another way of asking, "Are there brains in games?" and "If so, what else
can we say?"
In the experiment described below, we �nd a relationship between the heterogenous ability
to devote cognitive resources to a strategic interaction and behavior in the interaction. This
heterogeneity arises because we apply a di¤erential cognitive load on subjects who are playing
the game. In cognitive load experiments, subjects are directed to perform a memorization
task in parallel to making a choice in another domain. This additional memorization task
occupies cognitive resources, which cannot be devoted to deliberation about the choice. In
this sense, subjects under a larger cognitive load, can be thought to mimic the condition of
1See Camerer et. al. (1993), Johnson et. al. (2002), Crawford (2008), Costa-Gomes et. al. (2001) andCosta-Gomes and Crawford (2006).
2For instance, see Wang et. al. (2010) and Chen et. al. (2010).3For instance, see Coricelli and Nagel (2009).
2
having a diminished ability to reason.
Much is known about the behavior of subjects under a cognitive load. For instance, the
literature �nds that subjects under a larger cognitive load tend to be more impulsive and less
analytical. However, little is known about how the cognitive load a¤ects play in strategic
games.4
This experiment seeks to begin to clarify the relationship between cognitive load and
behavior in games. Further, due to the similarity between cognitive load and the diminished
ability to reason, the experiment seeks to sheds light on the relationship between intelligence
and behavior in games. One might be tempted to conclude that the diminished ability to
reason would generate obvious predictions; for instance that subjects under a larger cognitive
load will be more cooperative in the prisoner�s dilemma game. However, the predictions on
this front are far from obvious due to recent �ndings of a positive relationship between the
measure of intelligence and cooperation in the repeated prisoner�s dilemma game.5
In our experiment, we impose a cognitive load on subjects who are playing repeated multi-
player prisoner�s dilemma game. In each period, subjects are told to memorize a number. In
the low load treatment, this is a small number and therefore relatively easy to remember. In
the high load treatment, the number is large and therefore relatively di¢cult to remember.
The subjects then play a four-player prisoner�s dilemma game. After the subjects make their
choice in the game, they are asked to recall the number. As suggested above, subjects in
the low load condition are better able to commit cognitive resources in order to deliberate on
their action in the game.
Of course, the Subgame Perfect Nash Equilibrium of the �nitely repeated multi-player
prisoner�s dilemma game is for each player to select the uncooperative action in every period.
As with most experimental investigations of the prisoner�s dilemma game, we do not observe
4Researchers have also studied the e¤ects of the contraints on the complexity of strategies on outcomes inthe �nitely repeated prisoner�s dilemma game. For instance, see Neyman (1985, 1998). Also see Béal (2010)for a more recent reference. Our study can be thought to perform a similar exercise in the laborary.
5For instance, see Jones (2008).
3
this. We do �nd that the individual behavior of the subjects in the low load condition
converges to the Subgame Perfect Nash equilibrium prediction at a faster rate than those in
the high load treatment. However, we do not �nd the corresponding relationship involving
game outcomes. Speci�cally, there is no evidence of a signi�cantly di¤erent convergence of
game outcomes across treatments. A potential explanation for these two results, is our �nding
that subjects in the low load treatment are better able to adjust their strategy in response to
the outcome in the previous period than are those in the high load treatment. As a result,
they are better able to identify advantageous, temporary situations in which additional surplus
could be captured. Further, this agility o¤sets the trend towards playing uncooperatively.
These results combine to suggest that the availability of cognitive resources a¤ects strategic
behavior.
1.1 Related Literature
A typical cognitive load experiment would direct subjects to engage in a task which would
require mental e¤ort, in addition to making a choice in a di¤erent domain. One treatment
would be given a relatively easy task (low load treatment) and the other would be given a
relatively di¢cult task (high load treatment). The experimenter would then measure the
di¤erences in behavior between the treatments. This literature �nds that subjects under a
larger cognitive load tend to be more impulsive and less analytical because those in the high
load treatment are less able to devote cognitive resources to re�ect on their decision.
For instance, Shiv and Fedorikhin (1999) describe an experiment in which subjects were
given an option of eating an unhealthy cake or a healthy serving of fruit. The authors found
that the subjects were more likely to select the cake rather than the fruit when they were
under the high cognitive load.
Much is known about how the cognitive load a¤ects subjects in nonstrategic settings.
In addition to being more impulsive and less analytical (Hinson et. al., 2003) it has been
found that subjects under a cognitive load tend to be more risk averse and exhibit a higher
degree of time impatience (Benjamin et. al., 2006), make more mistakes (Ryvdal, 2007), have
4
less self control (Shiv and Fedorikhin, 1999; Ward and Mann, 2000), fail to process available
information (Gilbert et. al., 1988; Swann et. al., 1990), perform worse on gambling tasks
(Hinson et. al. 2002), are more susceptible to a social label (Cornelissen et. al., 2007), and
have di¤erent evaluations of the fairness of outcomes (Cornelissen et. al., 2011; van den Bos
et. al., 2006; Hauge et. al., 2009).
There is a literature which examines the relationship between the level of cognition and
play in games, without explicitly manipulating the cognitive load. For instance, Chen et. al.
(2009) measure the working memory of subjects and examine behavior in double auctions.
The authors �nd some evidence that subjects with a higher working memory perform better.
Devetag and Warglien (2003) �nd a relationship between the working memory capacity of a
subject and the congruence of play to that predicted by equilibrium. Also Bednar et. al.
(2010) describe an experiment in which subjects simultaneously play two distinct games with
di¤erent opponents.6 The authors �nd that behavior in a particular game is a¤ected by
corresponding paired game.
However, to our knowledge, there are only two papers which investigate the relationship
between the manipulation of cognitive load and behavior in games, Roch et. al. (2000) and
Cappelletti et. al. (2008). Roch et. al. (2000) found that subjects under the low cognitive
load condition requested more resources in a common resource game. However, in Roch et.
al. the subjects were not told the penalty if the sum of the group�s requests were more than
the amount to be divided. As a result, one cannot determine whether the cognitive load
manipulation implied di¤erences in strategic behavior or simply di¤erences in the regard for
norms which are not incentivized.
Cappelletti et. al. (2008) study behavior in the ultimatum game and vary the ability of
the subject to deliberate by manipulating time pressure and cognitive load. The authors
�nd that time pressure a¤ects the behavior of both proposer and responder. However, the
authors �nd that cognitive load does not a¤ect behavior as either a proposer or responder.
6Also see Savikhina and Sheremeta (2009).
5
In contrast, we �nd that cognitive load does a¤ect behavior in our setting. The di¤erence
in e¢cacy of the manipulation is likely due to the di¤erences in the incentivization of the
memorization task. We discuss this issue further below.
There is a recent interest in the relationship between intelligence and preferences.7 This
literature �nds a negative relationship between intelligence and both risk aversion and time
impatience. Note the similarity to the �ndings in the cognitive load literature. Therefore, to
the extent that manipulating cognitive load is analogous to manipulating the intelligence of
the subject, we now discuss the small literature on the relationship between intelligence and
behavior in games. Burnham et. al. (2009) demonstrate a relationship between a measure
of intelligence and strategic behavior in a beauty contest game. In other words, the authors
�nd that subjects with a higher measure of intelligence select actions which are closer to the
Nash Equilibrium of the beauty contest.
On the other hand, Jones (2008) �nds a relationship between cooperation in the repeated
prisoner�s dilemma and the average SAT scores at the university where the experiment was
conducted.8 In other words, Jones �nds a negative relationship between a measure of intelli-
gence and strategic behavior in the prisoner�s dilemma game.
Therefore, to the extent that an increased cognitive load simulates the e¤ect of a reduced
ability to reason, the two papers discussed above would seem to make opposite predictions in
our setting. Burnham et. al. (2009) would seem to predict that subjects in the high load
treatment will exhibit more cooperation in the prisoner�s dilemma game and Jones (2008)
would seem to predict that outcomes in the high load treatment will exhibit less cooperation
in the prisoner�s dilemma game. The experiment which we describe below will help distinguish
between these two predictions.
The answer, as it turns out, is a bit more subtle. Across all periods, we �nd little di¤erence
between either the individual behavior or the game outcomes of the subjects in the high and
7See Frederick (2005), Benjmin et. al. (2006), Burks et. al. (2008), Dohmen et. al. (2010), and Chen et.al. (2011).
8See Rydval and Ortmann (2004) for a similar result.
6
low load treatments. However, we �nd that the individual behavior of the low load subjects
converge to the SPNE behavior at a faster rate than high load subjects. We also �nd that
subjects in the low load treatment are better able to condition on past outcomes than are high
load subjects.
Finally, note that economists have become interested in studying the response times of
subjects.9 Research has found that longer response times are associated with more strategic
and less automatic reasoning. As we are manipulating the ability of the subjects to devote
cognitive resources to the problem, the response time will prove to be a useful measure in its
e¢cacy. In other words, we use the response time as a measure of the cognitive resources
devoted to the problem.
2 Method
A total of 60 subjects participated in the experiment. The subjects were graduate and
undergraduate students of Rutgers University-Camden. The experiment was conducted in
two sessions of 16, one session of 12, and two sessions of 8. The experiment was programmed
and conducted with the software z-Tree (Fischbacher, 2007).
Subjects were matched with three other subjects in which they were to play a repeated
prisoner�s dilemma game. The subjects were told that the group would remain �xed through-
out the experiment.10 The individual decision was to select X (the cooperative action) or Y
(the uncooperative action). Of the four subjects in the group, if x play X, and 4 � x play
Y then selecting X yields a payo¤ of 20x points whereas selecting Y yields 20x + 40. The
exchange rate was $1 for every 150 points. Additionally, the subjects were paid a $5 show-up
fee. While making a decision in the game, the subjects were provided with the payo¤s matrix
in two forms, which they were told are identical. See the appendix for the screen shown to
the subjects during their decision in the game.
9For instance, Brañas-Garza and Miller (2008), Piovesan and Wengström (2009), and Rubinstein (2007)10The instructions were given via power point slides. The slides, along with any experimental material, are
available from the corresponding author upon request.
7
Before play in each period, the subject was given 15 seconds in which to commit a number
to memory. The subjects were aware that they would be asked to recall the number after
their choice was made in the game. There were two cognitive load treatments: in the low
load treatment subjects were directed to memorize a 2 digit number, and in the high load
treatment subjects were directed to memorize a 7 digit number. There were 26 subjects in
the low load treatment and 34 in the high load treatment. The subjects were told that they
would only receive payment in periods in which they correctly recalled the number and that
they would receive nothing for the periods in which they incorrectly recalled their number.
After each period, subjects were given feedback regarding play in the game, however they
were given no information about their performance on the memorization task. Across all
treatments, the composition of 12 of the 15 groups were homogenous, in that they contained
only a single load treatment. However, there were 3 groups which were mixed in the sense
that that 2 subjects were in the low load treatment and 2 were in the high load treatment.
We refer to this group as mixed. The subjects were told nothing about the composition of
their group.
To summarize the timing in each period, subjects were given the number (7 digits of 2
digits), they made their choice in the game, they were asked to recall the number, and they
were given feedback on the game outcome but not the memorization task outcome. Each of
these stages were designed so that the subject would not proceed to the next stage until each
subject completes the prior stage. This procedure was repeated for 30 periods, with a new
number in each period. The amount earned by the subjects ranged from $6:47 to $20:20,
with a mean of $14:76.
At the conclusion of period 30, the subjects answered the following questions on a scale
of 1 to 7: Which featured into your decisions between X and Y , your prudent side or your
impulsive side (1 prudent, 7 impulsive)? How di¢cult was it for you to recall your numbers
(1 very di¢cult, 7 not very di¢cult)? How di¢cult was it for you to decide between X and
Y (1 very di¢cult, 7 not very di¢cult)? How distracting was the memorization task (1 very
8
distracting, 7 not very distracting)? and How many of the memorization tasks do you expect
that you correctly answered (1 none correct, 7 all correct)?
The z-Tree output speci�ed the time remaining when the Click to Proceed button was
pressed. In the output, there appeared instances of a time remaining of 99999. This output
seems to have occurred if the "Click to Proceed" button was pressed before the clock could
begin. In the stage in which the number was given, we recorded the 56 instances of an output
of 99999 as 16 because there were 15 seconds allotted. In the stage in which the game was
played, we recorded the 2 instances of an output of 99999 as 31 because there were 30 seconds
to decide.
2.1 Discussion of the Experimental Design
Before we get into the results, we discuss some issues related to the design of the experiment.
Although the cognitive load manipulation is rather common, to our knowledge, we are the
only example of a paper in which the manipulation is repeated. As a result, it was not obvious
to us whether we should balance the experiment so that each subject would undergo the high
and low loads an equal number of times. However, we decided to keep the subjects in a single
treatment throughout the experiment. In part, this decision was due to the results in Dewitte
et. al. (2005) which reports that the e¤ects of the cognitive load manipulation can be lasting.
Also note that we decided to use a 7 digit number as the high load manipulation because it is
standard in the literature and because Miller (1956) �nds that this tends to be near the limit
of the memory of subjects.11
Also note that the bulk of the cognitive load literature does not incentivise the memoriza-
tion task. To our knowledge, Benjamin et. al. (2006) and Cappelletti et. al. (2008) are the
only examples of experiments with such material incentives. Cappelletti et. al. (2008) pays
the subjects per correct digit. On the other hand, we pay the full amount earned in the game
for correct recall and we pay nothing for incorrect recall. However, like Cappelletti et. al.
(2008), we provide no feedback regarding the accuracy of the memorization task. We make
11Also, see Cowan (2001) for an updated view on the memory capacity literature.
9
these two design decisions in order to reduce the ability of the subject to strategically allocate
cognitive resources. In particular, we want to avoid providing an incentive for the subject to
seek an interior solution to the trade-o¤ of cognitive resources for the memorization task and
deliberation on the game.
We designed the experiment so that the subject would only enter the following stage when
all other players completed the current stage. This was done in order to mitigate the ability
of the subjects to strategically decide the timing of their decisions. In other words, due to
our experimental design, there was little incentive for the subjects in the low load condition to
quickly leave the stage where they are given their number because they would not immediately
proceed to the game stage. Additionally, the subjects in the high load condition could not
quickly make their decision in the prisoner�s dilemma game, in order to spill their number in
the memorization task, because they would not immediately proceed to the relevant stage.
We study the four-player prisoner�s dilemma12 because it has a few attractive features for
the purpose of examining the role of cognitive load in strategic games. The game is relatively
simple because the decision is binary and the game is linear. For the sake of simplicity, we did
not elect to use a more general public goods game. However, the four-player version requires
more thought than the two-player version because outcomes depend on the actions of three,
and not just one, opponent. Further, most of the subjects are familiar with the two-player
version and would likely import this prior experience into the experiment. For this reason,
we employed the four-player version.
3 Results
All �ve of the manipulation check questions demonstrated signi�cant di¤erences between the
high and low load treatments. Speci�cally, those in the high load treatment reported being
more impulsive (p = 0:038), having more di¢culty in recalling the number (p < 0:001), having
more di¢culty in deciding on an action in the game (p = 0:098), found the memorization task
12See Komorita et. al. (1980).
10
to be more distracting (p < 0:001), and expected to correctly recall the number with lower
precision (p = 0:005) than those in the low load treatment.13 Further, the subjects in the high
load spent a signi�cantly longer time committing the number to memory (M = 9:15, SD =
4:93) than did the subjects in the low load treatment (M = 1:19, SD = 2:20), t(1798) = 42:1,
p < 0:001.
We now begin the analysis of the individual behavior of the subjects. To do so, we perform
a series of logistic regressions with the choice in the game as the dependent variable. Here
a value of 1 indicates that the cooperative action (X) was selected and 0 indicates that that
the uncooperative action (Y ) was selected. We use a type dummy where 1 indicates the low
load treatment and 0 indicates the high load treatment. We use a dummy variable indicating
whether the group was mixed and therefore contained subjects from both the high and low
load treatments. Finally, we use a dummy variable indicating whether the memorization
task in that period was correct or incorrect. Note that the regressions below account for the
subject-speci�c �xed-e¤ects and each have n = 1800. See Table 1 for the results of these
regressions.
(1) (2) (3) (4) (5)
Period �0:0399��� � �0:0267��� �0:0267��� �0:0267���
(0:00627) (0:00801) (0:00801) (0:00801)Low Type � 0:0961 0:5849 0:6405 0:6517
(0:6654) (0:7110) (0:5829) (0:5838)Period-Low Type Interaction � � �0:0336��� �0:0336��� �0:0336���
(0:0130) (0:0130) (0:0130)Mixed � � � �0:1251 �0:1316
(2:5609) (2:5611)Mixed-Low Type Interaction � � � 0:5650 0:5632
(2:4751) (2:4754)Correct � � � � �0:0650
(0:1849)�2 Log L 2049:69 2091:41 2042:90 2042:90 2042:78LR �2 340:94��� 299:22��� 347:73��� 347:73��� 347:86���
Table 1: Fixed-e¤ects logistic regressions with a dependent variable of choice,
where *** indicates signi�cance at 0.01
13These are the results of a one-sided t-test between the subjects under high and low loads.
11
First, note that there is strong evidence of learning across periods. In every speci�ca-
tion involving the period, our results indicate that subjects played less cooperatively across
time. In other words, perhaps not surprisingly, we see convergence to the Subgame Perfect
Nash Equilibrium behavior. Although perhaps surprisingly, across periods, there was little
di¤erence between the choice of subjects in the high and low treatments. In each of the
four speci�cations involving the low type dummy, the variable does not achieve signi�cance.
However, the di¤erences between the treatments emerge when we account for time. The
actions of the subjects in the high load treatment converged to the equilibrium behavior at
faster rate than those in the low load treatment. This relationship continues to hold when
we account for the mixed nature of the groups or whether the subject correctly preformed
the memorization task in that period. Hence, there was convergence to the Subgame Perfect
Nash Equilibrium behavior for all types, however the convergence was faster for the low load
subjects.
One potential explanation for the faster convergence for the low load treatment is that
the high load treatment, due to the di¤erential di¢culty of the memorization task, expects
to earn less money than the low load treatment. As a result, the behavior of the high
load subjects converge to the low paying equilibrium prediction less quickly than the low
load subjects. However regression (5) provides evidence against this possibility: there is
no signi�cant relationship between choice and whether the subject correctly performed the
memorization task in that period.
The natural question is then, "Are the cognitive load treatments thinking di¤erently about
the game?" To answer this question, we analyze the response times of the subjects in selecting
an action in the game. We run the following three regressions with the time remaining as
the dependent variable In other words, the size of our dependent variable is increasing in
the speed of the decision. In each regression below, we account for �xed-e¤ects and have
n = 1800. The results are summarized in Table 2.
12
(1) (2) (3)
Period 0:247��� � 0:193���
(0:0148) (0:0196)Low Type � �3:90��� �5:85���
(1:51) (1:47)Period-Low Type Interaction � � 0:126���
(0:0297)R2 0:43 0:34 0:44
Table 2: Fixed-e¤ects linear regressions with a dependent variable of time
remaining in the game decision, where *** indicates signi�cance at 0.01
Again there appears to be a great deal of learning across periods. In both speci�cations
in which the period is included, there is a positive relationship between the period and the
speed of the decision. This suggests that as the experiment proceeded, the game decision
became more automatic and required fewer cognitive resources. The results of the regressions
involving type suggest that the subjects in the low load treatment re�ected on the decision
longer than did the high load subjects. Finally, the result of regression (3) suggests that the
low load subjects exhibited stronger learning across periods than did the subjects in the high
load treatment, as demonstrated by the positive interaction term.
Note the relationship between Table 2 and regressions (1)-(3) in Table 1. Indeed, in our
view, the results of Table 2 suggest an explanation for the results of Table 1. As previous
research has indicated, the response time is associated with more strategic and less automatic
reasoning. Therefore, the signi�cant, positive estimates of the period coe¢cients in Table 2
suggest that the subjects are becoming familiar with the game. This suggests an explanation
for the observation of the convergence to the Subgame Perfect Nash Equilibrium behavior.
The results of Table 2 also suggest that the low load subjects are becoming familiar with the
game at a faster rate than the high load subjects. Again this suggests an explanation for the
result that the individual behavior of subjects in the low load treatment were converging to
the Subgame Perfect behavior at a faster rate than the high load subjects.
It is also interesting to note that, unlike Table 1, Table 2 demonstrates a signi�cant rela-
tionship with the treatment dummy. In particular, we observe that subjects in the high load
13
treatment were responding faster than the low load subjects. A possible explanation for this
relationship is that the high load subjects exhibited a lower marginal bene�t of time thinking
about the game, because of the more di¢cult memorization task, which they must subse-
quently complete. Therefore, these di¤erences provide an explanation for the observation
that the high load subjects make their selection in the game at a faster rate.
Despite these di¤erences in individual behavior, the corresponding di¤erences for game
outcomes were not signi�cantly di¤erent across treatments. In particular, we do not �nd
the same di¤erential convergence of payo¤s as we did for choice. We perform the following
regressions, with the payo¤s earned in the game as the dependent variable. For the purposes
of the analysis below, we do not account for the accuracy in the memorization task. In other
words, in the regressions below, we use the payo¤s which would have been earned had the
memorization task been performed correctly. For this reason, we describe the dependent
variable to be provisional payo¤s. Note that up to this point, we what now describe as
provisional payo¤s, we referred to as game outcomes. We will henceforth use the term
provisional payo¤s. In each regression below, we account for �xed-e¤ects and have n = 1800.
These regressions are summarized in Table 3.
(1) (2) (3) (4) (5)
Period �0:313��� � �0:263��� �0:263��� �0:264���
(0:0489) (0:0650) (0:0650) (0:0650)Low Type � �3:33 �1:57 �1:57 �1:81
(4:69) (4:88) (4:88) (4:89)Period-Low Type Interaction � � �0:114 �0:114 �0:113
(0:0987) (0:0987) (0:0987)Mixed � � � 8:00� 8:23�
(4:64) (4:64)Mixed-Low Type Interaction � � � 7:33 7:29
(6:56) (6:56)Correct � � � � 1:38
(1:50)R2 0:17 0:15 0:17 0:17 0:17
Table 3: Fixed-e¤ects linear regressions with a dependent variable of provisional
payo¤s earned in the stage game, where * indicates signi�cance at 0.1, and ***
indicates signi�cance at 0.01
14
We �nd that the provisional payo¤s were decreasing across periods. This result is not
surprising because, as we found earlier, the individual behavior of the subjects was converging
to the Subgame Perfect behavior. We also �nd that the low type dummy variable is not
signi�cantly related to provisional payo¤s. Again, this is not surprising because we did
not �nd the analogous relationship between the low type dummy and individual behavior.
However, what is surprising is that, the provisional payo¤s of the low treatment subjects
do not converge at a rate di¤erent than that of the high load subjects. This is surprising
because, in the individual behavior regressions, there was a strong di¤erence in the convergence
of the high and low treatments. Also note that these relationships involving period, type,
and period-type interaction are robust to accounting for the mixed nature of the groups and
whether the subject correctly performed the memorization task in that period.
In both speci�cations involving the mixed group dummy, we see that there is a relationship
between the mixed group variable and payo¤s which is marginally signi�cant. This suggests
that subjects in mixed groups did better than subjects in homogenous groups. However,
within these mixed groups, there was no di¤erence between the high and low treatments. In
other words, conditional on being in a mixed group, those in the low load did not have a
signi�cantly di¤erent provisional payo¤ than those in the high treatment. Finally, we note
that the correct dummy is not signi�cantly related to provisional payo¤s.
Perhaps the most surprising aspect of the results to this point, is the strong signi�cance
of the period-type interaction in Tables 1 and 2, and its lack of signi�cance in Table 3. On
the one hand, we found that the individual behavior in the low load treatment converged to
the Subgame Perfect behavior faster than those in the high load treatments. These results
are found in Table 1. On the other hand, we found that the analogous result did not hold for
provisional payo¤s. Speci�cally, the provisional payo¤s of the high and low load treatments
did not converge at a di¤erent rate. These results are found in the Table 3. We now consider
a possible explanation for these two seemingly dissonant results: perhaps the low load subjects
were better able to condition on previous outcomes, and this extra agility o¤set the trend of
playing uncooperatively.
15
In order to explore this explanation, we run �xed-e¤ects logistic regressions with choice as
the dependent variable. As in the previous analysis, a 1 indicates that the cooperative action
was selected and 0 indicates that that the uncooperative action was selected. As we hope to
summarize the play in previous periods, we employ a variable which indicates the number of
other players in the group who played cooperatively in the previous period. In other words,
we compare the action selected in period t with the number of other group members who
played cooperatively in period t � 1. In the description below, we refer to this variable as
Lagged Number of Others Playing X. Note that this variable can range from 0 to 3. Another
possible measure of previous play is the change in cooperation between the previous period
and the period preceding that. In other words, we compare the action selected in period t
with the di¤erence in the number of other group members who played cooperatively in period
t� 1 and the number who played cooperatively in period t� 2. We refer to this variable as
Lagged Change in Others Playing X. Note that this variable can range from �3 to 3. Finally,
we include the three relevant interaction terms. In the regressions below, we account for
�xed-e¤ects. Due to the nature of the lagged variables, regression (1) has n = 1740, and
regressions (2) and (3) have n = 1680. The results are summarized in Table 4.
(1) (2) (3)
Low Type 0:157 0:526 �0:220(0:705) (11:7) (11:7)
Lagged Number of Others Playing X 0:0523 � �0:0733(0:0849) (0:125)
Interaction with Low Type 0:0677 � 0:431��
(0:133) (0:197)Lagged Change in Others Playing X � 0:0753 �0:0142
(0:0621) (0:110)Interaction with Low Type � �0:112 �0:317��
(0:0970) (0:137)Lagged Number of Others Playing X � � 0:0947�
-Lagged Change Interaction (0:0517)�2 Log L 1987:63 1894:62 1885:26LR �2 302:54��� 313:43��� 322:79���
Table 4: Fixed-e¤ects logistic regressions with a dependent variable of choice,
where * indicates signi�cance at 0.1, ** indicates signi�cance at 0.05, and ***
16
indicates signi�cance at 0.01
In regression (1) we do not observe a signi�cant relationship. As previous analysis
suggests, the treatment type is not related to choice. Also, we do not observe a relationship
between choice and the number of others playing cooperatively in the previous period. Further
there is not a signi�cant di¤erence between the sensitivity of the high load subjects to the
number of others playing cooperatively in the previous period and the sensitivity of the low
load subjects.
In regression (2), we observe a similar lack of signi�cance as that in regression (1). Again,
we observe that the type variable is not signi�cantly related to choice. We observe that the
lagged change in others playing cooperatively is not signi�cantly related to choice. Finally,
we do not observe a signi�cant relationship between the sensitivity of the high load subjects
to the change in the cooperation and the sensitivity of the low load subjects.
However, in regression (3) signi�cant relationships emerge. Again, the cognitive load type
variable is not signi�cant, nor are either of the measures of previous cooperation. But, we
do observe a di¤erential sensitivity to both measures of previous cooperation. The low load
types are more sensitive to the number of others playing cooperatively in the previous period
than the high load types. Additionally, the low load types are also more sensitive to the
change in the numbers of those playing cooperatively than the high load types.
Consider the signs of the variables indicating that the behavior of the low load subjects
was more sensitive than that of the high load subjects to previous outcomes. We note that the
interaction between the treatment and Number of Others Playing X is positive. This suggests
that low load subjects are more likely than high load subjects to cooperate in response to a
high level of cooperation in the previous period. We also note that interaction between the
treatment and the Change in Others Playing X is negative. This suggests that low load
subjects are more likely than high load subjects to play uncooperatively in response to an
increase in cooperation between the previous period and the period preceding the previous
period.
17
Although the lack of signi�cance in regressions (1) and (2) above, seems dissonant to the
signi�cance in regression (3), intuition on the matter is relatively straightforward. Behavior
is not exclusively a function of the level of cooperation in the previous period or exclusively
a function of the change in the cooperation, but it is a function of both. Consider a subject
making a decision regarding choice, where 2 of the 3 other subjects played cooperatively in the
previous period. By itself, the number of cooperators in the previous period has no context,
and is therefore not a su¢cient basis on which to make the choice. If the number of cooperators
rose from 1 to 2, the decision maker would regard that as di¤erent from the situation in which
the number of cooperators fell from 3 to 2. Therefore, signi�cant relationships only emerge
when we consider the level of cooperation and the change in cooperation.
To further analyze the role of type in the sensitivity of choice to previous outcomes, we
run the following �xed-e¤ects logistic regressions. In regression (1), we restrict to only the
high load subjects. In regression (2) we restrict to only the low load subjects. The results
are summarized in Table 5.
(1) (2)
Lagged Number of Others Playing X �0:0706 0:354��
(0:125) (0:154)Lagged Change in Others Playing X 0:0252 �0:385���
(0:123) (0:145)Lagged Number of Others Playing X 0:0639 0:138�
-Lagged Change Interaction (0:0677) (0:0802)�2 Log L 1128:28 756:487LR �2 126:12��� 197:078���
n 952 728
Table 5: Fixed-e¤ects logistic regressions with a dependent variable of choice,
where * indicates signi�cance at 0.1, ** indicates signi�cance at 0.05, and ***
indicates signi�cance at 0.01
The results of regression (1) suggest that neither the number of others playing coopera-
tively in the previous period, nor the lagged change in others playing cooperatively, nor their
interaction is signi�cantly related to choice for the high load types. By contrast, the results
of regression (2) indicate that each of the variables attains some level of signi�cance. In
18
particular, the number of others playing cooperatively is signi�cantly related to choice of the
low load subjects at the 0:05 level. Further, the lagged change in others playing cooperatively
is related to choice for the low load subjects at the 0:01 level. Together the results in Tables 4
and 5 suggest that the choice of the low load subjects was more sensitive to previous outcomes
than the choice of the high load subjects.
We now test the robustness of the result that the low load subjects were more sensitive to
previous play than the high load types. Although we �nd that the result is in general robust,
we also �nd that in one speci�cation, the signi�cance is greatly reduced. Here we perform
the identical analysis to that in summarized in Table 4. However, here we also account for
the mixed nature of the groups and the time e¤ects. In the regressions below, we account for
�xed-e¤ects and have n = 1680. We summarize these results in Table 6.
(1) (2) (3) (4)
Low Type �0:220 �0:503 �0:343 0:251(11:7) (0:626) (0:633) (0:714)
Lagged Number of Others Playing X �0:0733 �0:0733 �0:196 �0:157(0:125) (0:125) (0:128) (0:129)
Interaction with Low Type 0:431�� 0:431�� 0:397�� 0:281(0:197) (0:197) (0:199) (0:210)
Lagged Change in Others Playing X �0:0142 �0:0142 0:0786 0:0623(0:110) (0:101) (0:112) (0:112)
Interaction with Low Type �0:317�� �0:317�� �0:312�� �0:248�
(0:137) (0:137) (0:138) (0:142)Lagged Number of Others Playing X 0:0947� 0:0947� 0:0825 0:0764-Lagged Change Interaction (0:0517) (0:0517) (0:0521) (0:0522)Mixed � �2:00 �1:98 �2:18
(66:1) (65:6) (65:1)Mixed-Low Type Interaction � 2:40 2:38 2:59
(66:1) (65:6) (65:1)Period � � �0:0340��� �0:0236��
(0:00736) (0:00929)Period-Low Type Interaction � � � �0:0277�
(0:0153)�2 Log L 1885:26 1885:26 1863:54 1860:22LR �2 322:79��� 322:79��� 344:52��� 347:83���
Table 6: Fixed-e¤ects logistic regressions with a dependent variable of choice,
where * indicates signi�cance at 0.1, ** indicates signi�cance at 0.05, and ***
19
indicates signi�cance at 0.01
In order to facilitate the robustness check, regression (1) in Table 6 is identical to regres-
sion (3) in Table 4. Regression (2) accounts for the mixed nature of the groups, and the
di¤erential sensitivity of the low subjects to previous play remains signi�cant. Regression
(3) also accounts for the period of the decision, and again the di¤erential sensitivity of the
low subjects to previous play remains signi�cant. However, when we account for the mixed
nature of the groups, the period and the period-type interaction, as we do in regression (4),
we see that the di¤erential sensitivity of the low load types is diminished. In particular we see
that the di¤erential sensitivity to the number of others playing cooperatively in the previous
period is not signi�cant at any level. Further, the di¤erential sensitivity to the change in
cooperation is only signi�cant at the 0:1 level.
3.1 Discussion
In the experiment described above, we found that behavior of both high and low load subjects
in the multi-player prisoner�s dilemma converged to the Subgame Perfect behavior. However,
across all periods, we did not �nd a di¤erence in the behavior of the high and low treatments.
When we consider the time and the treatment then we note another signi�cant relationship:
the behavior of the low load subjects converged to the uncooperative Subgame Perfect pre-
diction at a faster rate than did that of the high load subjects. However, when we perform
the similar analysis, but with the provisional payo¤s, we note that there was no di¤erential
convergence of game outcomes for the types.
One potential explanation for these two seemingly incongruent results is that low load
subjects were better able to condition behavior on previous outcomes, and this agility o¤set
the general trend towards the uncooperative outcomes. In particular, we found evidence that
the low load subjects could, better than high load subjects, sustain cooperation when the level
of cooperation in the previous period was high. We also found evidence that the low load
subjects were more likely, than high load subjects, to play uncooperatively when there was an
increase in the level of cooperation between the previous period and the period preceding that.
20
In other words, it seems that the low load subjects were better able to identify advantageous,
temporary situations in which additional surplus could be captured.
So it seems that, while subjects in the high load treatments were more cooperative, and
this would seem to imply higher provisional payo¤s, this bene�t of cooperation seems to have
been o¤set by their reduced ability to condition actions on previous outcomes. The net result
of these two e¤ects, which work in opposite directions, results in no signi�cant di¤erences in
either the provisional payo¤s or the convergence rates of the provisional payo¤s.
4 Conclusion
So are there brains in games? And if so, what else can we say? Our results suggest a quali�ed
yes to the �rst question. Given our manipulation of the availability of cognitive resources in
our particular strategic environment, we found that di¤erences in cognitive resources imply
di¤erences in strategic behavior.
Regarding the second question, the answer is somewhat delicate. We found that subjects
within the low load treatment converged to the equilibrium prediction at a faster rate than
did those under the high load. However, we found no di¤erences in the convergence of the
payo¤s. To reconcile these two results, we note that the low load subjects were better able to
condition their play on previous outcomes. This agility of the low load types seems to allow
them to identify a temporary, advantageous situation and capture the available surplus. This
agility seems to o¤set the e¤ect on payo¤s trend of playing uncooperatively.
There seem to be two ways to slice the results of the experiment. On the one hand,
the reader who is not sympathetic to behavioral arguments, will point to the evidence of
the convergence towards the Subgame Perfect behavior of both cognitive load treatments.
Indeed, we found that subjects, even in the high load treatment, exhibited behavior which
converged to that predicted by the theory. This seems to support the claim that subjects,
even those with diminished cognitive resources, will eventually learn from their mistakes and
therefore intelligence is ultimately of limited concern in strategic settings. Further, the lack
21
of signi�cance of the treatment dummy variable in the results involving choice or provisional
payo¤s, could also be used to support the claim that the cognitive resources available to the
subject is of limited interest in a strategic setting.
On the other hand, the reader who is more sympathetic to behavioral arguments will note
that the di¤erences between the cognitive resources available to the subjects were directly
related to the di¤erences in the rate of the convergence to the equilibrium behavior. Indeed,
we found that the subjects in the low load treatment converged to the equilibrium behavior
at a faster rate than did the subjects in the high load treatment. Further, we found evidence
that the low load subjects were more sensitive than high load subjects to previous outcomes.
These results seem to o¤er support to the claim that the cognitive resources available to the
subject are of interest in strategic settings. Despite the position of the reader, we hope that
this experiment begins to clarify the role of cognitive resources in strategic settings.
The relationship between cognitive resources and play in games is also of interest to re-
searchers who study nonequilibrium models. In response to the mounting evidence that
subjects rarely play according to the equilibrium predictions, researchers turned their atten-
tion to nonequilibrium models which can account for errors made by subjects (McKelvey and
Palfrey, 1995) or hierarchical levels of thinking (Camerer et. al., 2004; Costa-Gomes, et. al.
2001). It would seem natural to conclude that the intelligence of the subject would be related
to either the errors committed or to the level of thinking employed by the subject. How-
ever, Georganas et. al. (2010) found that the mapping of measures of intelligence to the
hierarchical level of thinking varies across games. While there could be other reasons for this
negative result,14 evidence of this kind is crucial in supporting existing nonequilibrium models
or suggesting modi�cations to existing models. While the repeated nature of our present
experiment does not allow a clean comparison to this literature, our paper suggests that it
could be fruitful to investigate the relationship between the nonequilibrium models and the
intelligence of subjects, through the application of a di¤erential cognitive load.
14See Crawford et. al. (2010).
22
There remain several interesting and unanswered questions. For instance, it is unclear
how the results would be a¤ected by a game other than the multi-player prisoner�s dilemma.
For instance, it is unclear how the di¤erence in behavior of the treatments would be a¤ected
by an increase (for instance, a public goods game or auction) or a decrease (for instance, a
two-player prisoner�s dilemma) in the computational di¢culty of the game. We hope that
future work will examine the role of the complexity of the game.
Another unanswered question relates to the signi�cance of the incentives regarding the
memorization task. While our cognitive load manipulation was successful, and we found no
evidence of a relationship between choice and whether the memorization task was correct in
that period, it is possible that the subjects exhibited an income e¤ect. In other words, since
payment was only made when the memorization task was correct, and the memorization task
for the high load types was more di¢cult, it is possible that the subjects acted di¤erently as
a result of the �nancial incentives rather than the di¤erential cognitive resources. We also
hope that future work can address the a¤ect of our incentives on our results.
Finally, note that we only applied a cognitive load during the stage in which the subjects
selected an action in the game. We conjecture that our results would be strengthened if the
load was applied during both the game decision stage and the feedback stage. However, only
future work could test this conjecture.
23
5 Appendix
The screen during the game decision:
24
6 References
Béal, Sylvain (2010): "Perceptron versus automaton in the �nitely repeated prisoner�s dilemma,"Theory and Decision, 69: 183�204.
Bednar, Jenna, Chen, Yan, Liu, Tracy Xiao, and Page, Scott (2010): "Behavioral Spilloversand Cognitive Load in Multiple Games: An Experimental Study," working paper Universityof Michigan.
Benjamin, Daniel, Brown, Sebastian, and Shapiro, Jesse (2006): "Who is �Behavioral?�Cognitive Ability and Anomalous Preferences," working paper Harvard and Chicago.
Brañas-Garza, Pablo and Miller, Luis (2008): "Instinctive Response in the UltimatumGame," working paper Universidad de Grenada and Oxford University.
Burks, Stephen V., Carpenter, Je¤rey P., Götte, Lorenz and Rustichini, Aldo (2009):"Cognitive Skills Explain Economic Preferences, Strategic Behavior, and Job Attachment,"Proceedings of the National Academy of Sciences, 106(19): 7745�7750.
Burnham, Terence C., Cesarini, David, Johannesson, Magnus, Lichtenstein, Paul and Wal-lace, Björn (2009): "Higher cognitive ability is associated with lower entries in a p-beautycontest," Journal of Economic Behavior and Organization, 72: 171�175.
Camerer, Colin F., Ho, Teck-Hua, and Chong, Juin Kuan (2004): "A Cognitive HierarchyModel of Games," Quarterly Journal of Economics, 119(3): 861�898.
Camerer, Colin F., Johnson, Eric J., Rymon, Talia and Sen, Sankar (1993) "Cognition andFraming in Sequential Bargaining for Gains and Losses." In Frontiers of Game Theory, ed.Ken Binmore, Alan Kirman, and Piero Tani. MIT Press, Cambridge, MA, 27�47.
Cappelletti, Dominique, Güth, Werner, and Ploner, Matteo (2008): "Being of two minds:an ultimatum experiment investigating a¤ective processes," working paper, Jena EconomicResearch Papers 2008-048.
Chen, Chia-Ching, Chiu, I-Ming, Smith, John, and Yamada, Tetsuji (2011): "Too smartto be sel�sh? Measures of intelligence, social preferences, and consistency," working paperRutgers University-Camden.
Chen, Chun-Ting, Huang, Chen-Ying and Wang, Joseph Tao-yi (2010): "A Window ofCognition: Eyetracking the Reasoning Process in Spatial Beauty Contest Games," workingpaper National Taiwan University.
Chen, Shu-Heng, Tai, Chung-Ching, and Yang, Lei-Xieng (2009): "Agent-Based Mod-eling of Cognitive Double Auction Market Experiments," working paper National ChengchiUniversity.
25
Coricelli, Giorgio and Nagel, Rosemarie (2009): "Neural correlates of depth of strategicreasoning in medial prefrontal cortex," Proceeings of the National Acadamy of Science, 106(23), 9163�9168.
Cornelissen, Gert , Dewitte, Siegfried, Warlop, Luk, and Yzerbyt, Vincent (2007): "What-ever people say I am, that�s what I am: Social labeling as a social marketing tool," Interna-tional Journal of Research in Marketing, 24: 278�288.
Cornelissen, Gert, Dewitte, Siegfried, and Warlop, Luk (2011): "Are Social Value Orienta-tions expressed automatically? Decision making in the dictator game," Personality and SocialPsychology Bulletin, forthcoming.
Costa-Gomes, Miguel A., Crawford, Vincent P. and Broseta, Bruno (2001): "Cognition andBehavior in Normal-Form Games: An Experimental Study," Econometrica, 69(5): 1193�1235.
Costa-Gomes, Miguel A., and Crawford, Vincent P. (2006): "Cognition and Behavior inTwo-Person Guessing Games: An Experimental Study," American Economic Review, 96(5):1737�1768.
Cowan, Nelson (2001): "The magical number 4 in short-term memory: A reconsiderationof mental storage capacity," Behavioral and Brain Sciences, 24(1): 87�185.
Crawford, Vincent P. (2008): "Look-ups as the Windows of the Strategic Soul.� In TheFoundations of Positive and Normative Economics: A Handbook, ed. Andrew Caplin andAndrew Schotter. Oxford University Press: Oxford, 249�280.
Crawford, Vincent P., Costa-Gomes, Miguel A. and Iriberri, Nagore (2010): "StrategicThinking," working paper UCSD.
Devetag, Giovanna and Warglien, Massimo (2003): "Games and phone numbers: Do short-term memory bounds a¤ect strategic behavior?" Journal of Economic Psychology, 24: 189�202.
Dewitte, Siegfried, Pandelaere, Mario, Briers, Barbara, and Warlop, Luk (2005): "Cogni-tive load has negative after e¤ects on consumer decision making," working paper KatholiekeUniversiteit Leuven, Belgium.
Dohmen, Thomas, Falk, Armin, Hu¤man, David, and Sunde, Uwe (2010): "Are RiskAversion and Impatience Related to Cognitive Ability?" American Economic Review, 100(3):1238�1260.
Fischbacher, Urs (2007): "z-Tree: Zurich Toolbox for Ready-made Economic Experi-ments," Experimental Economics, 10(2): 171�178.
Frederick, Shane (2005): "Cognitive Re�ection and Decision Making," Journal of EconomicPerspectives, 19(4): 25�42.
Georganas, Sotiris, Healy, Paul J., and Weber, Roberto A. (2010): "On the Persistence ofStrategic Sophistication," working paper Ohio State University.
26
Gilbert, Daniel T., Pelham, Brett W., and Krull, Douglas S. (1988): "On Cognitive Busy-ness: When Person Perceivers Meet Persons Perceived," Journal of Personality and SocialPsychology, 54 (5): 733�740.
Hauge, Karen Evelyn, Brekke, Kjell Arne, Johansson, Lars-Olof, Johansson-Stenman, Olof,and Svedsäter, Henrik (2009): "Are Social Preferences Skin Deep? Dictators under CognitiveLoad," working paper University of Gothenburg.
Hinson, John M., Jameson, Tina L., and Whitney, Paul (2002): "Somatic markers, workingmemory, and decision making," Cognitive, A¤ective, and Behavioral Neuroscience, 2 (4): 341�353.
Hinson, John M., Jameson, Tina L., and Whitney, Paul (2003): "Impulsive DecisionMaking and Working Memory," Journal of Experimental Psychology: Learning, Memory, andCognition, 29(2): 298�306.
Johnson, Eric J., Camerer, Colin, Sen, Sankar, and Rymon, Talia (2002): "DetectingFailures of Backward Induction Monitoring Information Search in Sequential Bargaining,"Journal of Economic Theory, 104: 16�47.
Jones, Garett (2008): "Are smarter groups more cooperative? Evidence from prisoner�sdilemma experiments, 1959�2003," Journal of Economic Behavior and Organization, 68: 489�497.
Komorita, S. S., Sweeney, James, and Kravitz, David A. (1980): "Cooperative Choicein the N-Person Dilemma Situation," Journal of Personality and Social Psychology, 38 (3):504�516.
McKelvey, Richard S. and Palfrey, Thomas R. (1995): "Quantal Response Equilibria forNormal-Form Games," Games and Economic Behavior, 10(1): 6�38.
Miller, George (1956): "The magical number seven, plus or minus two: Some limits on ourcapacity for processing information," Psychological Review, 63 (2): 343�355.
Nagel, Rosemarie (1995): "Unraveling in Guessing Games: An Experimental Study,"American Economic Review, 85(5): 1313�1326.
Neyman, Abraham(1985): "Bounded complexity justi�es the cooperation in the �nitelyrepeated prisoner�s dilemma," Economic Letters, 19, 227�229.
Neyman, Abraham(1998): "Finitely repeated games with �nite automata," Mathematicsof Operations Research, 23, 513�552.
Piovesan, Marco and Wengström, Erik (2009): "Fast or Fair? A Study of Response Times,"Economics Letters, 105: 193�196.
Roch, Sylvia G., Lane, John A. S., Samuelson, Charles D., Allison, Scott T. and Dent,Jennifer L. (2000): "Cognitive Load and the Equality Heuristic: A Two-Stage Model of
27
Resource Overconsumption in Small Groups," Organizational Behavior and Human DecisionProcesses, 83(2): 185�212.
Rubinstein, Ariel (2007): "Instinctive and Cognitive Reasoning: A Study of ResponseTimes," Economic Journal, 117: 1243�1259.
Rydval, Ondrej (2007): "Financial Incentives and Cognitive Abilities: Evidence from aForecasting Task with Varying Cognitive Load," working paper Max Planck Institute.
Rydval, Ondrej and Ortmann, Andreas (2004): "How �nancial incentives and cognitiveabilities a¤ect task performance in laboratory settings: an illustration," Economics Letters,85: 315�320.
Savikhina, Anya and Sheremeta, Roman M. (2009): "Simultaneous Decision-Making inCompetitive and Cooperative Environments," working paper Purdue University and ChapmanUniversity.
Shiv, Baba and Fedorikhin, Alexander (1999): "Heart and Mind in Con�ict: The Interplayof A¤ect and Cognition in Consumer Decision Making," Journal of Consumer Research, 26(3):278�292.
Stahl, Dale and Wilson, Paul (1994): "Experimental Evidence on Players� Models of OtherPlayers," Journal of Economic Behavior and Organization, 25: 309�327.
Stahl, Dale, and Wilson, Paul (1995): "On Players� Models of Other Players: Theory andExperimental Evidence," Games and Economic Behavior, 10: 218�254.
Swann, William B., Hixon, Gregory, Stein-Seroussi, Alan, and Gilbert, Daniel T. (1990):"The Fleeting Gleam of Praise: Cognitive Processes Underlying Behavioral Reactions to Self-Relevant Feedback," Journal of Personality and Social Psychology, 59 (1): 17�26.
Van den Bos, Kees , Peters, Susanne L., Bobocel, D. Ramona, and Ybema, Jan Fekke(2006): "On preferences and doing the right thing: Satisfaction with advantageous inequitywhen cognitive processing is limited," Journal of Experimental Social Psychology, 42: 273�289.
Wang, Joseph Tao-yi, Spezio, Michael and Camerer, Colin F. (2010): "Pinocchio�s Pupil:Using Eyetracking and Pupil Dilation To Understand Truth Telling and Deception in Sender-Receiver Games," American Economic Review, 100(3), 984�1007.
Ward, Andrew and Mann, Traci (2000): "Don�t Mind If I Do: Disinhibited Eating UnderCognitive Load," Journal of Personality and Social Psychology, 78 (4): 753�763.
28